2M+ Downloads · 9.3K Stars
The open-source framework has over 2M downloads, 9.3K stars on GitHub and 600+ releases in the last 12 months.
PraisonAI helps banks, insurers and hospitals run AI agents on their own computers, even fully offline, so their private data never leaves the building.
Most AI agents send your data to someone else's servers. Banks, insurers and hospitals can't do that with customer or patient records, so their AI work waits on approvals that take months.
A rented model sees every record you send it.
Compliance teams want proof before they trust AI with their data.
OpenAI takes no new self-serve fine-tuning jobs after 6 January 2027.
Each task moves to your own AI only after it passes a test. If quality slips, it moves back automatically, and every step is recorded for regulators.
The open-source framework is live today. The handover is in early access, and we are inviting our first design partners.
Records how each task is done today.
Your own AI takes an exam made from real past work.
A task moves to your own AI only when it passes.
If quality drops, the task moves back automatically.
Every handover leaves a record your regulator can check.
Your own AI is an open model, such as Llama, Qwen or Gemma, on your own computers. Use it as it is, or train it on your own past work. Either way, it takes over a task only after it passes an exam made from real past work it never studied.
The open-source framework is live today. The handover shown here is in early access.
Agents run on your own computers with your own models, such as Llama, Qwen and Gemma through Ollama or vLLM. They keep working with no internet connection.
Customer and transaction data stays under your own controls.
Claims and policy records stay on your own systems.
Patient records never leave the building.
Connect models and tools to turn an idea into an agent that can act.
Build multi-agent workflows for research, planning and execution.
Bring memory and your own knowledge into the workflow with RAG.
This installs PraisonAI. When it finishes, run praisonai setup to add a provider key. With a local model such as Ollama, no key is needed.
curl -fsSL https://raw.githubusercontent.com/MervinPraison/PraisonAI/main/install.sh | shpip install praisonaiagentsnpm install praisonaipip install praisonai
praisonai agents.yamlexport OPENAI_API_KEY="your-key"The repository opens on this agent. It takes a question, does the research, and returns a table. This is the first job. A larger one is an Agent Team or an Agent Flow.
Tell the agent it is a senior analyst.
Ask for the research, and for a table back.
A finished table, ready for the next draft or the next agent.
from praisonai import Agent
agent = Agent(instructions="You are a senior data analyst.")
agent.start("Analyze the top 3 tech trends of 2026 and format as a markdown table.")An agent can answer in the chat people already use, and a scheduler can keep the job running. The dashboard stays on your machine.
Telegram, Discord, Slack, and WhatsApp. Add a channel from the dashboard.
A support reply can use what the agent remembers and what it has read.
An agent scheduler runs the job on a clock.
Each course teaches one idea. The example is on the page, and you run the agent yourself. No account is required.
Set up locally when needed, pick a track, and run one example.
Read short definitions for agent, RAG, MCP, and other terms used on the hub.
Build one agent, add tools and memory, run a team and handoffs, then deploy — each core step adds production-style constraints.
Chain steps, route, parallelise, and loop — each pattern adds tighter output rules.
Search a note, answer only from retrieved lines, and catch wrong-section answers.
Save facts, recall by keyword, and stack checks before you trust memory in production.
Author skills, load them on demand, and keep skills separate from tools.
Move through build, run, observe, and govern using the right University track for each phase.
Stack guardrails and judges — fail bad text on purpose before you trust scores.
Follow a role-based route across existing University tracks in a sensible order.
Run the same agent in a chat app, then schedule one repeating job.
Attach a tool server and call one MCP tool end-to-end.
Define one agent in a file, validate it, then run it with the CLI.
Put one agent behind a web address, call it, then run it in a container.
Turn on telemetry, read token cost, and send one agent run to a trace backend.
Block unsafe output, approve tool runs, and keep secrets out of source.
Pick a model, set a key safely, limit tokens in prompts, and run agent questions with names and format rules.
Chat with agents in the browser, manage sessions and jobs, and read usage without new scripts.
See three end-to-end stories and which tracks to open next.
Use an honest checklist until a platform certificate exists.
Plan, build, and ship one small agent end-to-end on Praison.
We’re inviting our first design partners. Talk to us about early access. The framework stays MIT licensed.
The open-source framework has over 2M downloads, 9.3K stars on GitHub and 600+ releases in the last 12 months.